EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1689 (emitting agent: issue-orchestrator).

Goal

In both Step 10d TG file-grain hit pipelines, normalize each leg's absolute tree root to a common token and exclude pytest warnings-summary lines from the file-grain grep; node-grain remains the failure arbiter.

Workflow gap

Bug observed: The Step 10d TG file-grain compare subtracts full lines including absolute tree prefixes (gated leg in the worktree, baseline at the repo root), so payload-path-bearing warning lines never cancel and passing branch-new tests' warnings have no baseline twin - innocent merges block.
Why it is a workflow gap: The gate's own documented semantics say pre-existing trunk red never blocks and node-grain is the failure arbiter, but the file-grain comm -23 (grep -F -f payload-paths | sed line-number-blank) is structurally incapable of cancelling absolute-path-bearing lines across the two trees, and a pytest warnings-summary line from a PASSING test is not a failure signal at all. On #1689 (2026-07-30) this cost 3 full gate cycles (~50 min): gate run 3 had node-grain NEW = 0 and lint legs clean, yet blocked on 5 file-grain lines — 4 were prefix-twins of the baseline-red-on-main test_issue825_mlp_batched_parity warning lines, 1 was a warnings-summary line from the branch's own passing test_issue1689_fit_ladder.py (branch-new test absent at root). The orchestrator had to re-derive the verdict manually from the persisted leg outputs.
Confidence (emitter): high
verified-at-filing: grep -n 'tg-gated-hits\|tg-baseline-hits' .claude/skills/issue/SKILL.md → 4 hits in 2 blocks (the shared form (i)/(ii) gate block ~L11337 and the form (iii) surgical block ~L13018) (2026-07-30)

Proposed change (candidate diff sketch — refine in planning)

In BOTH TG hit pipelines (shared gate block + surgical block), amend the per-leg hit extraction:

  for leg in baseline gated; do
    grep -F -f /tmp/issue-<N>-tg-files.txt "/tmp/issue-<N>-tg-$leg.txt" \
      | grep -vE '^E +assert ' \
+     | grep -vE ':: [A-Za-z]+Warning: ' \
      | sed -E 's/at line [0-9]+/at line N/g; s/:[0-9]+:/::/g; s/:[0-9]+([^0-9]|$)/:N\1/g' \
+     | sed -e "s|$WT|<TREE>|g" -e "s|$REPO_ROOT|<TREE>|g" \
      | sort -u \
      > "/tmp/issue-<N>-tg-$leg-hits.txt" || true
  done

(Surgical block: both legs run at the root, so only the warnings-exclusion line is load-bearing there; add the prefix normalization anyway for uniformity.) Rationale: warnings are not failures (node-grain + the scan tests' dedicated per-file evidence lines remain the block surface); the tree-prefix token makes the same line from the two trees cancel under comm -23.

Scope / surfaces

Primary target: .claude/skills/issue/SKILL.md
Grep the workflow surface for the pattern before editing (grep -rn 'tg-gated-hits' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan. Check whether any pin test asserts the current pipeline shape (grep -rl 'tg-baseline-hits' tests/) and update it in the same diff.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).
The node-grain arm and the scan-test per-file evidence lines stay byte-unchanged — this fix ONLY removes the two structural false-positive classes from the file-grain arm.

Provenance

workflow_fix_target: .claude/skills/issue/SKILL.md
fingerprint: 27381c19626d

target_file: .claude/skills/issue/SKILL.md bug_observed: The Step 10d TG file-grain compare subtracts full lines including absolute tree prefixes (gated leg in the worktree, baseline at the repo root), so payload-path-bearing warning lines never cancel and passing branch-new tests' warnings have no baseline twin - innocent merges block why_workflow_gap: The gate's documented semantics say pre-existing trunk red never blocks and node-grain is the failure arbiter, but the file-grain comm -23 is structurally incapable of cancelling absolute-path-bearing lines across the two trees; #1689 burned 3 gate cycles on a node-grain-clean payload proposed_change: In both Step 10d TG file-grain hit pipelines, normalize each leg's absolute tree root to a common token and exclude pytest warnings-summary lines from the file-grain grep; node-grain remains the failure arbiter diff_sketch: |

| grep -vE ':: [A-Za-z]+Warning: ' \
| sed -e "s| $WT|<TREE>|g" -e "s|$ REPO_ROOT||g"
(inserted into the per-leg hit pipeline in the shared gate block ~L11330 and the surgical block ~L13010) confidence: high related_task: #1689